Promised Slides from Hotchips conference

AlgebraicRing · Aug 25, 2005

I've just scanned what I think might be relevant. I'd make these pictures, but there is a limit to the number that can be opened.

Acert93 · Aug 25, 2005

Whoa, some good information there.

I only had time to read one of them (looked at the Xbox one to see if they had any new XeCPU info... nothing really new).

BUT it seemed to be a very even handed presentation of the problems the designers faced via customer requests (i.e. MS said "IBM, here are our hurdles") and how the hardware was designed to handle this. Without argueing numbers it just presented the hardware for what it was without overstating things and just highlighted the unique features of the design.

Some good stuff there!

Guden Oden · Aug 25, 2005

Thank yous!

This was some great stuffs you posted up there. The bus and I/O of cell and the x360 slides were the ones I could make at least some sense of, the first batch of cell slides were very technical and just gave me a headache.

MAN. EIB 300GB/s aggregate throughput. That's some serious $#!&! Almost unbelievable, and then I look down on the floor where my now ancient NES sits with its on-die 2 kilobyte SRAM and a processor clock speed of what is it, 2.3MHz or something like that? Jeeeeeez!

Now it's giga-this and giga-that. We've almost completely dumped the mega prefix altogether, heh!

Crazy when you think about it.

Shifty Geezer · Aug 25, 2005

I think this side should be cause for concern...

Scarey!

http://i12.photobucket.com/albums/a210/AlgebraicRing/talk_1_slide_9.jpg

Acert93 · Aug 25, 2005

I like the Intel "big fat chip" approach better:

Stuff in => CPU => Stuff out.

Really... that slide has to be a "ok, 15 minute break so you can clean your panties, after that we will show you the real slide" hahaha

Shifty Geezer · Aug 25, 2005

It was slide 9. Hopefully IBM broke them into this gently. Somehow. Maybe they laced the coffee with liquer? Honestly though trying to keep tabs of all that and all those interactions is surely enough to send coders running screaming into the hills!

Jawed · Aug 25, 2005

The "streaming" support in XB360 seems fairly extensive. For example the way that read streams don't consume L2 space, with the data being loaded directly into L1.

Then the non-coherent data produced by a program consuming the streamed data is effectively made coherent in L2, and read out as a stream by the GPU.

Slide 9 shows this in action. Seems pretty smart to me.

Jawed

Guden Oden · Aug 25, 2005

Jawed,
I agree. x360 design looks very intelligent and well-thought-out. It might be a bit too intelligent tho, the engineers who designed it were probably smarter than the coders that will program it! On the other hand. SNES had some pretty wacky functionality too that couldn't be leveraged unless you really went off the deep end and immersed yourself in the hardware, with all the HDMA funkyness and stuff. Super Aleste was one of those titles that really exploited HDMA to the fullest. Sprites everywhere with hardly any slowdown.

AlgebraicRing · Aug 25, 2005

Jawed said:
The "streaming" support in XB360 seems fairly extensive. For example the way that read streams don't consume L2 space, with the data being loaded directly into L1.

Then the non-coherent data produced by a program consuming the streamed data is effectively made coherent in L2, and read out as a stream by the GPU.

Slide 9 shows this in action. Seems pretty smart to me.

Jawed

Yep, slide 9 seems to be the magic. It is going to be interesting to see which console comes out on top in terms of graphics. The XBOX 360 design looks pretty slick according to the slide, but what's it cost in terms of programmer time and hand tuning to make sure the pipes are all brimming with data and flowing in the right directions (not crashing into each other)?

At this point I don't know that the XBOX 360 is going to be any easier to program for. It looks like there could be a lot of concurrency/parallelism (data scheduling) that needs to be specifically coded for.

Jawed · Aug 25, 2005

I simply don't know enough about the algorithms used to produce higher-order surfaces, or the kind of workload involved in procedurally generating geometry (e.g. trees), but it seems to me that these could be two good uses for this streaming architecture.

I suppose the same also applies to lighting calculations generating texture data for consumption by the GPU.

At the same time, you have to ask yourself when tasks like these are better performed by the GPU instead? One core of Xenon has the computing power of, oh, let's say about 8 of Xenos's pipes. And in streaming type applications, a GPU is the bee's knees.

So I suppose it boils down to a comparison of the memory access patterns of the (streaming) code. But if you have a Xenon core spare, you might as well use it as an adjunct.

Jawed

MrFloopy · Aug 25, 2005

Scarey!

Just shows how bad Sony is at effective developer presentation. (check some Sony dev session presentations).

Honestly though trying to keep tabs of all that and all those interactions is surely enough to send coders running screaming into the hills!

This graph has nothing to do with practical programming or interactions. It's marketing.

On the MS side, not really anything new and interesting for you guys except for this:

8 outstanding load/prefetches per core

this is sweet sweet music.

rendezvous · Aug 25, 2005

MrFloopy said:
Just shows how bad Sony is at effective developer presentation. (check some Sony dev session presentations).

The messy graph slide doesn't come from Sony though.

And i don't think that it has that much to do with programming either.
I belive that it just shows dependencies between design decisions made while developing the Cell architecture. The developers shouldn't be too worried about it.
(Disclaimer: It is hard to be sure what they really mean without the audio presentation though)

AlgebraicRing · Aug 25, 2005

rendezvous said:
The messy graph slide doesn't come from Sony though.

And i don't think that it has that much to do with programming either.
I belive that it just shows dependencies between design decisions made while developing the Cell architecture. The developers shouldn't be too worried about it.
(Disclaimer: It is hard to be sure what they really mean without the audio presentation though)

The slide was definitely referring to the design decisions. The left side were what they wanted to accomplish, the right side was what they ended up with (I think). I think the purpose of the slide is to really point out how essential the nodes with lots of arrows going in and out are. I guess it's supposed to make you feel better about the fact that you'll have to do a lot of "Static Scheduling".

I agree with Mr. Floopy, Sony isn't making the programming for the CELL sound easy and XBOX 360 had a nice presentation and a concrete programming model to offer. But I wonder just how easy it will be to pull off that programming model. There's gonna be gotchyas on both systems.

dukmahsik · Aug 25, 2005

xCPU just keeps sounding better and better for games

Urian · Aug 25, 2005

I am not scared when I see the diagrams of the Cell and the Xenon.

The terror for the programmers isnÂ´t the architecture, are the programming tools.

darkblu · Aug 25, 2005

xbox360 slide 13

main memory fb/zb => unpredictable performance.

oh, the irony... makes you wonder what exactly they were thinking when they were designing the original xbox..

aaaaa00 · Aug 25, 2005

darkblu said:
xbox360 slide 13
oh, the irony... makes you wonder what exactly they were thinking when they were designing the original xbox..

Just speculating but:

"We need to ship something fast! What can we put together that will be fast enough to beat PS2 by a significant margin in most scenarios and still be fairly cheap, and can go from final project OK to design, test, and production in 18 months?"

darkblu · Aug 25, 2005

aaaaa00 said:
Just speculating but:

"We need to ship something fast! What can we put together that will be fast enough to beat PS2 by a significant margin in most scenarios and still be fairly cheap, and can go from final project OK to design, test, and production in 18 months?"

yeah, it beat the ps2 in general computing. totally wiped the floor with it. somebody had/has a fairly amusing idea of 'most scenarios' at redmond. and it was fairly cheap. maybe compared to a pc, alright. gee, xbox being overly cheap must be the reason why ms abandoned it in record time. but the final question that really bugs me to no end is exactly what stopped ms from coming up with a _real_ console design? what was the magic reason for releasing that green mistake in the timeframe they did? ..some honcho at some corporate meeting must have had just a tad too much of self-confidence, would be my speculation.

scooby_dooby · Aug 26, 2005

you should listen to Ed Fries Kikizo interview.

Some head honchos including Gates were on a business retreat and were kicking around the idea of getting into the console business.

At the same time, there were 2 console projects being developed internally. One was the XBOX, named after DirectX that it was based on. Ed, who was head of MGS, lent his weight to the XBOX project as he believed it was a much more gamer-driven project.

So what happened was there was some desire from above to enter the console market, and desire from below and the two sort of met in the middle, and XBOX was released. So the XBOX was designed internally by PC devs, it was built to be the most powerful console out there adn it was, without question. So what's the problem?

Ed goes on to say that when they began work on the xbox's successor, it became too many cooks in the kitchen. Everyone from different divisions was trying to control/influence the project because they were seeing what an important trojan horse it could be.

He quit in 2003(i think) because of this, he lost control, and that's pretty much the reason why MS has changed directions so drastically I think. At one time this was sort of a side project for MS, by gamers for gamers, Fries was all about creating the most powerful console on the market, but now it's a huge priority, all about profit and comprimises, it's kinda gay if you ask me.

The whole no video streaming with windows XP, the 'core' package, or ridiculous peripheral pricing are all examples of this...

richardpfeil · Aug 26, 2005

I noticed something on the #5 360 slide. Video Out is show as a box inside the GPU, directly connected to a box labeled Analog Chip. Doesn't Dave's Xenos expose state that the chip has no video out? I thought it wrote the buffer to main memory. Also, this diagram seems to be saying that the lack of HDMI output in inherent in the chip, thus there can be no new cable created to support it.

Promised Slides from Hotchips conference

AlgebraicRing

Acert93

Artist formerly known as Acert93

Guden Oden

Senior Member

Shifty Geezer

uber-Troll!

Acert93

Artist formerly known as Acert93

Shifty Geezer

uber-Troll!

Jawed

Guden Oden

Senior Member

AlgebraicRing

Jawed

MrFloopy

rendezvous

AlgebraicRing

dukmahsik

Urian

darkblu

aaaaa00

darkblu

scooby_dooby

richardpfeil

Similar threads